Leveraging user-generated content with latent semantic indexing for cold-start playlist continuation

Music recommender systems, which are indispensable components of online music streaming services, aim to provide better music listening experience to their users by providing personalization in line with users' musical preferences. An important requirement in achieving this goal is the automatic continuation of playlists in a coherent manner. A music recommender system should identify the characteristics of a playlist, create a user profile based on these characteristics, and recommend music suitable for this profile. This requirement, which is quite challenging due to the huge size of music catalogs, becomes even more complex for new playlists that do not yet contain any tracks. In such cold-start cases, the use of additional information to reveal possible music themes in playlist is essential. In this dissertation, playlist naming tendencies of users consuming music via the Internet are analyzed. Furthermore, it is investigated how the common attitude in playlist organization can be utilized in music recommender systems. Based on the findings obtained from a large dataset containing one million different playlists, a novel recommendation approach using latent semantic indexing is proposed to alleviate the cold-start problem in automatic playlist continuation. The experimental studies and results show that the proposed method can produce recommendations with higher accuracy than the state-of-the-art approaches in the literature for the addressed cold-start problem. The statistically significant improvements on a large, real-world dataset reveal that latent semantic indexing can effectively discover hidden relationships between user-generated titles and music preference.